Humans are just like GenAI. Inputs matter.

Recently, I’ve started learning about generative AI (GenAI), such as ChatGPT and Google Bard. The power and potential of these large language model (LLM) tools are just beginning to be understood and unlocked by those in the mainstream. However, in evaluating the capabilities of these systems, I found myself making an unexpected connection. Seeking information and conversing with humans follows the same rules and limitations as working with ChatGPT.

A core issue with GenAI is that their responses are only as good as the information on which they were trained. As they continue to evolve and "learn," each input influences their future responses. Similarly, a child effectively begins as a blank slate. We teach them by reading, showing them videos, and talking to them. They begin to recall shapes, colors, animals, etc. Gradually, as more source material is provided, they make increasingly complex connections. However, the range of their responses is inherently limited by the inputs of their experiences.

Does this really ever change? Every day, nearly eight billion people wake up and spend their day gaining new experiences. We incorporate these inputs into our perspective on the world. This shapes how we respond to future prompts. Countless factors structure the inputs we receive. We’re each inherently limited by our location and how we interact with the world. Having consistent daily experiences and doing so with people with similar inputs reinforces our perspective. With a limited range of inputs, trends within that scope support our knowledge. We find a local optimal response, unaware of the potential global optimal response, by not exploring other potential areas of our theoretical and physical maps. Just like our LLM counterparts, inputs create consistency in our response. Correlation begins to imply causation because that matches our experience. In this article, I’ll explore several factors that contribute to this limitation and make a logical argument for DEI efforts and their impact related to how we think about artificial and human intelligence.

Seeking Positive Reinforcement: Confirmation Bias

Several fundamental psychological forces influence the quality of information we feed our internal LLMs. Humans are innately biased to seek information that agrees with our beliefs. Confirmation Bias is the first critical factor impacting how we train our brains. Rather than taking in a diverse collection of information and drawing a conclusion based on all available data, we often form our hypothesis first, then search for information that supports that concept, rejecting data that proves us wrong. It's more comfortable and assuring to be correct. Admitting we were wrong, even to ourselves, feels stressful. On the mild side, confirmation bias is reading reviews for a product we already bought, focusing on positive experiences to validate our decision, and rejecting the opinion of those with a negative perspective. At a greater extreme, it’s adopting a conspiracy theory and searching the depths of the internet to find more evidence and a community that agrees with this newly held perspective. With each additional confirming source, we’re reinforcing connections, giving us (false?) confidence to recite the information with greater conviction.

Learning takes effort. If integrating new information feels easy and comfortable, chances are you're not actually learning something but reinforcing what you already believe. Integrating new inputs should be challenging, and we are often forced to confront a series of questions. Why did we think that was true? What previous data conflicts with these new inputs? Which is really correct, or is neither accurate, and the "truth" is somewhere in the middle, or different entirely? Confronting confirmation bias takes guts, energy, and commitment, but that’s actually how learning happens. When asked to respond to a new prompt, you’ll have a more robust data set to pull on and the ability to engage in an actual dialogue rather than merely recite a series of facts that only align with a single perspective.

Indirect Experience: Learning through Empathy

The diversity of our inputs significantly impacts how we are trained to respond to and view a given situation. For many, we're born into a specific community, initially adopting the perspectives of our parents/guardians and close acquaintances, trained to see our situation as "normal." It's all we know, like trying to explain water to a fish. “That's just the way it is. Obviously, the way we see things is how everyone else sees them.” Only by expanding our experience basis directly and indirectly can we begin to appreciate the differences between our experience and that of others. For some, it's the opportunity to travel and immerse in a completely different environment from our upbringing. The earlier this happens, the better to shape the direct inputs to our experiences and appreciate both subtle and more significant differences. However, we all can experience the world indirectly as well by seeking others’ perspectives.

Speaking from my experience, I grew up in a more rural area north of Pittsburgh, PA. My first immersion into the experience of others was at my elementary school, Big Knob, where I learned from other local children. In 5th grade, we were thrust into the bigger world with the merger of the other elementary school in my district – kids that lived “in town” and went to Conway. Wow! What a different life they led! They could walk to the local strip mall! What must that experience have been like for them?

Fast forwarding to today, I’ve been lucky enough to live in four different states in different quadrants of America, each bringing their unique experience. I’ve worked closely with those from other continents, hearing their input histories growing up in Argentina, India, and China. I began to understand the differences in how they saw the world, and it dramatically eclipsed the perceived differences between Big Knob and Conway.

Empathy is a critical characteristic to expand our internal LLM, growing beyond our direct experiences and integrating the perspective of others to help understand situations and why others react differently than we do. This takes effort and an openness to delay our internal reactions to a situation and see things through someone else's eyes. We must actively listen to gain this valuable input, pausing our internal response mechanism to be in a learning phase. If you respond to a situation with "Well, everyone I know believes X" or "voted for Y," consider that this still might not be the "correct" response. The limiting factor may be the scope of pooled data. It takes work, but you’ll be more informed and better equipped to assess responses with an openness to gain indirect experience inputs.

Confidence in Limited Data

Humans have another significant hindrance in learning – the propensity to misjudge our level of expertise. While many LLMs now include phrases highlighting the boundaries of the information presented, human conversations often lack such a disclaimer. Usually, the more limited the foundational knowledge, the greater the confidence and conviction with which it is communicated. The Dunning-Krueger Effect classified this obstacle of querying humans for information. As our personal LLM gains a small, fundamental amount of information, the wiser we feel, often eager to impart our prowess to others. Just like the confirmation bias above, digging further into a subject uncomfortably confronts us with the limitations of our understanding. At this stage, we can just sit at the peak of (false) confidence or genuinely gain awareness of the boundaries of our newly discovered “expertise.”

I've personally fallen victim to this phenomenon and have been on the long road to recovery since. Back in high school, I fell in love with chemistry. It came relatively easy, and I was pretty good at it, spawning the next two decades of my career. I was more than willing to answer any chemistry question in high school. I gained increasing and validated confidence given my limited pool of queries, prompting me to establish my first Hotmail account – chem_god. Yep, the hubris was strong with this one. As I continued learning at college, I remained at the top of my class. Yet, I began to recognize the boundaries of my learning and how much further the realm of chemistry extended beyond my data inputs. Continuing to pursue my doctorate, I was more directly confronted with my limitations and the expertise of others, putting my perceived mastery in appropriate perspective. Even today, after obtaining an advanced degree and working for years in related industries, I need to be reminded of my expertise as I have a greater awareness of my limitations rather than my knowledge. The shift from initially learning to becoming an expert changes perspective from what you know to an enhanced recognition of how much there is yet to learn.

We must maintain perspective as we receive new inputs and fight the bias of false confidence in our limited skills. Indeed, given the breadth of human scholarship, it’s impossible to attain mastery in every subject, and we need to define our value in greater apprehension in a specific area. While we each draw the line at which we need to establish our expertise, be mindful of the limitations of your comprehension. Seek to recognize the stage of proficiency of others you query for information. And feel free to include your personal disclaimer.

We’re each only as good as our inputs

The human mind remains a fantastic biological invention. Our ability to receive inputs, make connections, share, and expand human knowledge is incredible. As we seek to create machines that replicate aspects of our brain's capability, remember that the boundaries of these new systems also reflect some inherent limitations of our cognitive systems. If you only feed your mind junk or false information, recognize the reduced scope and value of your outputs. Remember the failure that was Microsoft’s Tay. If you train yourself with inaccuracies and conspiracies, don’t be surprised when that is how you begin to respond. Do your best to employ empathy to assess your inputs from others’ perspectives, and while you may not perceive the issue, recognize that in itself is restricted by your input data.

We each need to seek to overcome these challenges through a willingness to confront our biases. Just like using ChatGPT, recognize the role diverse inputs play in shaping our responses and the responses of others. Confront these limitations by overcoming the discomfort involved with the intake of conflicting information. Challenge yourself to seek information beyond that which reinforces already-held beliefs. Listen with an openness to gain value from the experiences of others to expand our empathetic data set. And provide disclaimers, recognizing the boundaries of our knowledge.

Continue growing your internal LLM by seeking good inputs to establish new, more robust connections. Help contribute to our information evolution in what you learn and how you enable others to understand. Good luck in your continued growth as human intelligence, and may we constantly strive to be better than our artificial counterparts.